AI model challenge in Accelerator based Nuclear and Particle Physics
Critical Opportunity:
- Experiments at future accelerators such as the Electron Ion Collider (EIC) will employ data streaming systems that preserve virtually all the data such that a fully unbiased study of the data sample, together with accelerator data, can be made.
- These massive datasets, rich in the complex physics embedded within, are an ideal basis to draw on the techniques of the AI LLM revolution to bring a transformative change in deriving scientific insights from experimental HENP data.
- We have the opportunity to build a fully cognizant facility from accelerator to detector and analysis.
- BNL is ideal for this work as the only US laboratory hosting multiple user facilities across different science domains, and is home to the only collider in the US
- RHIC, US ATLAS and the EIC comprise some of the largest HENP datasets available today and in the future.
Expected Impacts:
- Experimental HENP complexes such as the EIC are multi billion dollar enterprises that warrant application of the most sophisticated techniques to maximize their discovery potential, teasing out the greatest possible science return
- In economic impact,
- AI-driven efficiency improvements will yield the same data in much less time, reducing operations and energy costs.
- Can reduce the compute and storage intensive demands of HENP analysis, saving compute and energy costs.
- Cost saving example: Improving the signal to background in recorded EIC data by 10% through the use of AI (intelligent DAQ) would lead to a $300k/year saving for archiving media.
- Techniques developed for accelerators should be readily transferable to medical and industrial applications
- An ideal training ground for the nation’s AI workforce, ensuring continued US leadership in this critical new technology
Required R&D:
- AI dataset integration of accelerator, experiment and calibration/QA systems, e.g. for real-time optimization of luminosity and background; optimization in both directions between the machine and the experiment
- Create a HENP LLM (HENP data elements replacing ‘words’), and develop training and feature extraction techniques
- Develop techniques to enhance trustworthiness in building and using AI
Timeline:
- Near term: 1-3 years
- Complete R&D on using AI to draw trusted, quantified inferences from real data
- Develop a prototype HENP LLM with current data (e.g. ATLAS, RHIC) and first generation feature extraction tools
- Mid term: 3-5 years
- Begin work on an integrated accelerator - experiment dataset and its AI instrumentation for a cognizant facility from accelerator to detector to analysis, targeting the EIC
- Long term: 5-10 years
- Documented analyses employing the HENP LLM (internal notes and/or peer reviewed publications)
- Commission and deploy AI for EIC in parallel with machine and detector installation, commissioning and datataking